Amazon Web Services (AWS) this week unfurled a raft of capabilities to streamline the building and deployment of artificial intelligence (AI) models using either its Amazon SageMaker or Amazon Bedrock services.
In addition, AWS has added an integration capability for databases running on its cloud service that eliminates the need to employ extract, transform and load (ETL) tools to aggregate data that, for example, could be used to train an AI model.
Announced at the AWS re:Invent 2023 conference, additions such as Amazon SageMaker HyperPod, a pre-configured instance of the platform, promise to reduce the time needed to train foundation models by up to 40% using purpose-built infrastructure while Amazon SageMaker Inference reduces foundation model deployment costs on average by 50% on average latency by 20% using processors optimized by AWS.
Amazon SageMaker Clarify, meanwhile, provides a set of tools to evaluate and select foundation models based on their parameters that support responsible use of AI.
Additionally, AWS has added an Amazon SageMaker Canvas tool to accelerate data preparation using natural-language instructions to invoke AI models to automate data aggregation.
At the same time, AWS is making additional foundational models available via the Amazon Bedrock service. In addition, there are also now model evaluation tools for evaluating models based on use cases.
AWS has also added a Knowledge Bases for Amazon Bedrock that makes use of vector databases to enable IT teams to extend the capabilities of a large language model (LLM) using proprietary data for Amazon Bedrock to automate various tasks.
The cloud service provider has also added Guardrails for Amazon Bedrock to enable policies to be applied to how generative AI applications are employed.
Each organization will need to determine how best to go about building AI models, but Amazon Bedrock, via a single application programming interface (API), provides a simpler way to get started says Kumar Chelliappa, general manager for AI and machine learning services for AWS.
In contrast, Amazon SageMaker is designed to appeal to data science teams that need a platform to streamline the customization of AI models, he adds.
Regardless of approach, organizations will need to exercise some care when selecting AI models based on their use case, notes Chelliappa. Each foundational AI model is going to be subject to different levels of potential hallucinations based on how it was trained, he says.
In addition, based on how many parameters an AI model supports, costs will vary, adds Chelliappa. In many use cases, organizations will not need the latest most expensive AI model to drive their application, he adds.
As organizations pervasively employ AI models, costs are going to rise. Most AI models used to build applications are billed for using tokens. Each input and output requires a token, so at a minimum, every time an AI model is invoked it requires two tokens. There is no standard pricing for tokens, so each organization will need to carefully negotiate the total cost of accessing AI models.
As such, organizations would need to, despite whatever enthusiasm exists for all things AI, focus their efforts on the handful of projects that have the most potential to deliver a meaningful return based on the high level of investment required.
